This thesis addresses the problems of biometrics : how a persons identity could be determined or validated by using some physical or behavioral characteristic. Biometry is one of the main research topics in the field of pattern recognition due to its impact on several applications in security and human-machine interaction environments. Several works focus on the improvement of the features extracted in the particular system being presented (face, fingerprint or speech recognition among others), or the metrics used to compare such features, in this work the classification stage is particularly tackled.A statistical approach is presented based on a well-known a-contrario validation strategy. Techniques based on such framework have been widely used in the fields of image processing and computer vision for the detection and matching of visual features. In this work, the method ability to detect outliers/inliers is exploited to detect when two compared biometric samples correspond to the same person. This method is adapted and applied to each of the usual biometric tasks.First, it is applied to the task of biometric verification, modeling it as a two- class classification problem. The introduced strategy was validated using different datasets and compared against other state-of-the-art commonly used classification methods. Findings of this work have been presented at the 2014 International Conference on Pattern Recognition Applications and Methods (ICPRAM-2014), by applying the framework to the face recognition problem in particular. An extension of the conference article has been published as a journal article. In this thesis, the presented strategy is reviewed with an experimental evaluation done in several bigger datasets.Secondly, the a-contrario framework is applied to the identification task. The method is used to validate the confidence of an identification system outputs. What is normally called in the literature as System Response Reliability (SRR). Such problem has been thoroughly studied lately, the key advantages of using such control are analyzed and discussed. The obtained performance is validated on multiple datasets by comparing with other state-of-the-art approaches. This work has been presented on the 2016 International Conference of the Biometrics Special Interest Group (BIOSIG-2016).Finally, the framework is applied to biometric fusion. The key differences in such scenario and the corresponding proposed framework adaptations are analyzed. The proposed technique is evaluated in both artificially generated as real-scenario datasets. The performance is compared against other state-of-the-art statistically fusion strategies